Drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning

Abstract

The invention relates to the field of bioinformatics, in particular to a drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning. The method comprises the following steps: S1, constructing a three-layer heterogeneous network based on existing data; s2, constructing an optimization model based on the three-layer heterogeneous network; s3, solving an objective function of the constructed model by using a block coordinate descent algorithm; and S4, predicting the potential correlation between the drug and the pathway based on the solving result of the objective function. According to the method, firstly, a drug and pathway association data set is constructed, and a data basis is provided for a subsequent association prediction task; related data of genes are introduced into a drug and pathway association prediction problem, and a three-layer heterogeneous network is constructed, so that data diversity and information richness can be increased; by constructing a regular optimization model, information in a heterogeneous network can be effectively mined to complete a prediction task. According to the method, the calculation efficiency is high, the prior information of drug-pathway association is utilized, the prediction accuracy is improved, and a negative sample is not needed.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting drug and pathway associations based on heterogeneous network cross-layer dependency reasoning. Background Art

[0002] Traditional drug development methods usually follow the "one drug, one target" strategy. However, the pathological process of complex diseases usually involves the interaction between many functionally related biomolecules, which are located in certain disease-related pathways. In addition, drugs usually work by affecting related signaling pathways rather than just targeting a single target. Many biological experiments have demonstrated various associations between drugs and pathways. For example, in one study, a patient with medulloblastoma used the drug GDC-0449, and the experimental results showed that GDC-0449 could inhibit the Hedgehog pathway, resulting in rapid tumor regression.

[0003] Drug-pathway association refers to the fact that drugs regulate the expression of genes in pathways by targeting one or more genes, thereby affecting the activity of pathways. In addition, drug-pathway associations can be divided into positive correlation and negative correlation according to the mechanism of action of drugs on pathways, corresponding to drugs activating or inhibiting pathways by increasing or decreasing gene expression. Drug-pathway associations provide more physiological or functional information for discovering chemical compounds that can treat complex diseases. Therefore, identifying drug-pathway associations is a key task to accelerate drug discovery and development. However, the number of known drug-pathway associations is limited, and discovering these associations through experimental methods is time-consuming and laborious. With the continuous development of experimental technologies such as genomics, proteomics, and metabolomics, the acquisition of drug sensitivity data and high-throughput transcriptional data has become more convenient. These data provide valuable information for discovering drug-pathway associations. Therefore, it is particularly important to establish effective computational methods to mine potential drug-pathway associations. Reliable inference results can not only promote experimental verification, but also greatly save human and financial resources.

[0004] At present, researchers have proposed and constructed some machine learning models for drug and pathway association prediction. These prediction models can be roughly divided into three categories: prediction models based on Bayesian sparse factors, prediction models based on matrix decomposition, and other machine learning models. In the Bayesian sparse factor method and matrix decomposition method, the gene transcription spectrum or drug sensitivity spectrum data of different human cell lines are used to infer the potential association between drugs and pathways. In the Bayesian sparse factor method, the association prediction problem is usually described by a linear model, and the collapsed Gibbs sampling algorithm is used for Bayesian inference. Representative prediction models include iFad and FacPad. The matrix decomposition rule converts the problem into a regularized optimization model. The feature matrix is ​​obtained by solving the objective function, and then the drug and pathway association prediction is performed. Representative prediction models include iPaD, L 2,1 -iPaD,L1L 2,1 -iPaD and IGMF. In addition, other machine learning-based methods integrate various data such as the chemical structure of drugs and gene expression related to pathways as features to represent drugs and pathways. Then, using training samples, different classifiers are used to train prediction models for drug and pathway association prediction. Representative prediction models include RGRF and LSA-PU-KNN.

[0005] In the prediction model based on Bayesian sparse factors, there are many parameters that need to be estimated, and the requirements for computing resources are relatively high. In the prediction model based on matrix decomposition, the core of these methods is to solve the target association matrix, but the association matrix can only reflect the association between drugs and pathways, and cannot reflect the probability of the association between drugs and pathways. In addition, the differences between the existing representative prediction model methods are small, mainly reflected in the penalty function used for the regularization term. For example, L 2,1 -iPad uses L 2,1 The norm penalty replaces the L1 norm penalty, and L1L 2,1 -iPaD combines the L1 norm and L 2,1The penalty terms of the L1 norm are added as regularization terms. In addition, IGMF combines L1 norm penalty and graph regularization. Existing models based on matrix decomposition do not fully utilize the prior information of drug-pathway associations when predicting potential drug-pathway associations, which may reduce the prediction accuracy to a certain extent. Other machine learning models also have some limitations. For example, some drugs or pathways have more associated samples in the training samples, which may cause the trained prediction model to have prediction bias problems. In addition, in supervised machine learning models, positive and negative samples are required to construct training sample sets. However, since irrelevant drug-pathway pairs are difficult to collect, it is difficult to obtain negative samples. In addition, in the field of drug and pathway association prediction, there is less prior information on drug and pathway associations, which will have an adverse effect on the accuracy of the prediction model at the data level. Some existing prediction models have not made some methodological and technical improvements to this problem. Summary of the invention

[0006] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0007] In view of the above problems and / or the problems existing in the prior art, the present invention is proposed.

[0008] The object of the present invention is to provide a method for predicting drug and pathway associations based on heterogeneous network cross-layer dependency reasoning, comprising the following steps:

[0009] S1. Build a three-layer heterogeneous network based on existing data;

[0010] S2. Construct an optimization model based on a three-layer heterogeneous network;

[0011] S3. Solve the objective function of the constructed model using a block coordinate descent algorithm;

[0012] S4. Predict the potential association between drugs and pathways based on the results of solving the objective function.

[0013] Further, step S1 specifically includes S11 and S12;

[0014] S11. Construct drug layer network, pathway layer network and gene layer network respectively; S12. Construct three-layer heterogeneous network: based on the known drug and pathway association and pathway and gene association data, connect the nodes in the pathway network layer with the nodes in the drug network layer and the nodes in the gene network layer respectively to obtain a three-layer heterogeneous network.

[0015] Furthermore, in step S11, the nodes in the drug network layer network represent different drugs. If the two drugs corresponding to the two nodes are similar, the two nodes are connected by an edge, and the weight of the edge is the similarity value of the two drugs. The nodes in the pathway network layer network represent different pathways. If the two pathways corresponding to the two nodes are similar, the two nodes are connected by an edge, and the weight of the edge is the similarity value of the two pathways. The nodes in the gene network layer network represent different genes. If the two genes corresponding to the two nodes are similar, the two nodes are connected by an edge, and the weight of the edge is the similarity value of the two genes.

[0016] Furthermore, in step S11, a drug similarity matrix S is first constructed based on the existing data. d , Pathway Similarity Matrix S p and the gene similarity matrix S g ; Based on the similarity matrix S d , Pathway Similarity Matrix S p and the gene similarity matrix S g Get similarity value;

[0017] In some embodiments, the existing data include but are not limited to NCI-60 data sets and CCLE data sets. Based on NCI-60 data and CCLE data, different three-layer heterogeneous networks can be constructed. Although the heterogeneous networks are different, the drug and pathway association prediction method of heterogeneous network cross-layer dependency reasoning proposed in the present invention is universal.

[0018] Furthermore, in step S11, the calculation of the drug similarity value is based on the drug similarity matrix S d ; The drug similarity matrix S d The construction is based on the drug sensitivity spectrum matrix Y1 in the cell line, and the Pearson correlation coefficient is used to calculate and construct the drug similarity matrix. The specific calculation method is as follows:

[0019]

[0020] Where nc represents the number of cell lines in the data set, and the absolute value of the calculated result is taken, that is, S d (i,j)=|S d (i,j)|, as the final similarity value between drugs i and j.

[0021] Furthermore, in step S11, the calculation of the similarity value of the path is based on the similarity matrix S of the path. p ; The similarity matrix S of the pathway p The construction is based on the pathway and gene association matrix L1, and the Gaussian spectral kernel method is used to calculate the pathway similarity matrix. The specific calculation method is as follows:

[0022] S p (i,j)=exp(-α p ||L1(i;,)-L1(j;,)|| 2 ) (4)

[0023]

[0024] Where np represents the number of pathways in the dataset, S p (i,j) represents the similarity value between paths i and j.

[0025] Furthermore, in step S11, the calculation of the gene similarity value is based on the gene similarity matrix S g ;

[0026] The similarity matrix S of the genes g The construction is based on the gene expression spectrum matrix Y2 in the cell line and uses the Pearson correlation coefficient to calculate the gene similarity matrix. The specific calculation method is as follows:

[0027]

[0028] Where ng is the number of genes in the dataset; S g (i,j) represents the similarity value between genes i and j.

[0029] Further, step S2. constructs an optimization model based on the three-layer heterogeneous network: uses a regular optimization model to model the drug and pathway association prediction problem; uses the model to mine the potential feature representation of each layer of the heterogeneous network; the objective function of the model is:

[0030]

[0031] Where i, j represent the i-th and j-th layers in the three-layer heterogeneous network; i, j = 1, 2, 3 represent the drug layer, pathway layer, and gene layer, respectively; g = 3 represents the number of layers in the network;

[0032] G(i,j)=1 means that there is an association relationship between the nodes in the i-th layer and the j-th layer;

[0033] A i,j Represents the association matrix between nodes in the i-th layer and the j-th layer;

[0034] S i is the similarity matrix between the nodes in the i-th layer network;

[0035] is a weight matrix, n i and n j Represents the number of nodes in layer i and layer j respectively;

[0036] Matrix T i is the matrix S i The diagonal matrix

[0037] F i is the low-rank feature matrix of the i-th layer network.

[0038] Furthermore, in step S3, by solving the objective function, a block coordinate descent algorithm is used to solve the local optimal solution of the objective function; and the potential feature matrix of each layer of the three-layer heterogeneous network is obtained. By solving the objective function of the above model, the potential feature matrix of each layer of the three-layer heterogeneous network can be obtained. Since the objective function of the model is non-convex, a block coordinate descent algorithm is used to solve the local optimal solution of the objective function. The update rule of the feature matrix of each layer of the heterogeneous network is as follows:

[0039]

[0040] in When A i,j (u,v)>0 when I i,j (u,v)=1,A i,j (u,v)=0 when I i,j (u,v)=0; the final F can be obtained by iteratively updating the formula i (i=1,2,3); iterative convergence condition is: reaching the maximum number of iterations t=100 or three characteristic matrices F i The bi-norm of the corresponding difference values ​​of (i=1,2,3) in two iterations is lower than the threshold ξ=10 -8 .

[0041] Further, in step S4, the potential association between drugs and pathways is predicted based on the result of solving the objective function: after solving the potential feature matrix F of the drug layer, pathway layer, and gene layer, i (i=1,2,3), for a given drug i and pathway j, the association score AS(i,j) between them is defined as:

[0042] AS(i,j)=F1(i,:)·F2(j,:) T (11).

[0043] The two drug and pathway association datasets constructed in the present invention are the NCI-60 dataset and the CCLE dataset.

[0044] In one embodiment, the present invention calculates and constructs a drug similarity matrix based on the NCI-60 data set. Pathway Similarity Matrix and the gene similarity matrix First, we use the Pearson correlation coefficient to calculate and construct the drug similarity matrix based on the drug sensitivity spectrum matrix Y1 in the cell line. The specific calculation method is as follows:

[0045]

[0046]

[0047] Where nc=57 represents the number of cell lines in the NCI-60 dataset. Since the value range of the Pearson correlation coefficient is [-1,1], we take the absolute value of the calculated result, that is, As the final similarity value between drugs i and j.

[0048] Then, we calculated the pathway similarity matrix based on the pathway and gene association matrix L1 using the Gaussian spectral kernel method as follows:

[0049]

[0050] Among them, np=58 represents the number of pathways in the NCI-60 dataset, Represents the similarity value between paths i and j.

[0051] Then, we calculated the gene similarity matrix based on the gene expression profile matrix Y2 in the cell line using the Pearson correlation coefficient as follows:

[0052]

[0053] Wherein, ng=1863 represents the number of genes in the NCI-60 dataset.

[0054] Based on the CCLE dataset, the drug similarity matrix can be constructed using the same method Pathway Similarity Matrix and the gene similarity matrix

[0055] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are at least:

[0056] (1) The method provided by the present invention first constructs a drug and pathway association data set to provide a data basis for subsequent association prediction tasks; introducing gene-related data into the drug and pathway association prediction problem and constructing a three-layer heterogeneous network can increase data diversity and information richness; the drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning proposed by the present invention can effectively mine information in the heterogeneous network to complete the prediction task by constructing a regularized optimization model.

[0057] (2) Compared with the prediction model of Bayesian sparse factors, the present invention has high computational efficiency and low demand for computing resources. When predicting potential drug-pathway associations, existing models based on matrix decomposition do not fully utilize the prior information of drug-pathway associations, which will reduce the prediction accuracy to a certain extent; the technical solution of the present invention fully utilizes the prior information of drug-pathway associations to improve the prediction accuracy. The technical solution of the present invention does not require negative samples, which overcomes the difficulty of obtaining negative samples. (In supervised machine learning models, constructing a training sample set requires positive samples and negative samples. Since irrelevant drug-pathway pairs are difficult to collect, it is difficult to obtain negative samples.) BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application.

[0059] Figure 1 : Flowchart of the drug and pathway association prediction method based on cross-layer dependency reasoning in heterogeneous networks. DETAILED DESCRIPTION

[0060] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the embodiments of the specification. These embodiments are only used to illustrate the present application and are not used to limit the scope of the present application. Those skilled in the art can make similar generalizations without violating the connotation of the present invention, and the present invention is not limited by the specific embodiments disclosed below.

[0061] 5-fold Cross-Validation: The implementation effect of the technical solution of the present invention is evaluated by a 5-fold cross-validation method. The 5-fold cross-validation randomly divides the data set into five parts, uses one subset as the test set each time, and performs five independent validations.

[0062] iPaD method: The iPaD method (integrative Pathway Analysis and Drug discovery) is a method that uses matrix decomposition to predict the association between drugs and pathways.

[0063] AUC index: AUC (Area Under the Curve) is an important indicator for evaluating the performance of a binary classification model, and is usually used to evaluate the performance of the model under different classification thresholds. It is the area under the ROC curve (Receiver Operating Characteristic curve). The larger the AUC value, the more effectively the model can distinguish between positive and negative classes. Here, based on the results of five-fold cross validation, we calculate the AUC value to evaluate the prediction accuracy of the method proposed in the present invention.

[0064] 1. The present invention first constructs two drug and pathway association datasets, namely the NCI-60 dataset and the CCLE dataset.

[0065] NCI-60 dataset: contains 101 drugs and 58 pathways, as well as 30 known drug-pathway association pairs. We use the association matrix A1∈R 101×58 To record the association information between drugs and pathways, if drug i and pathway j are known association pairs, then the element A1(i,j) in the matrix = 1, otherwise A1(i,j) = 0. For 101 drugs, we obtained their drug sensitivity spectrum data in 57 cell lines, using the matrix Y1∈R 57×101 To represent this data, the matrix element Y1(i,j) represents the GI of drug j in cell line i 50 (GrowthInhibition 50) value. GI 50 It is an indicator used to measure the inhibitory effect of a compound on cell proliferation, indicating the concentration of the compound required to inhibit cell growth by 50%. For 58 pathways, we obtained the association data between them and 1863 genes, using the matrix L1∈R 58×1863 To record the association data, if pathway i and gene j are associated, the element L1(i,j) in the matrix is ​​1, otherwise L1(i,j) is 0, and the number of associations between pathways and genes is 4273. For these 1863 genes, we also obtained their expression profile data, using the matrix Y2∈R 57×1863 to represent the expression profile of a gene in a cell line, where element Y2(i,j) represents the expression level of gene j in cell line i.

[0066] CCLE dataset: contains 22 drugs, 58 pathways, and 195 pairs of drug-pathway associations. Similarly, we use the association matrix A2∈R 22×58 To represent the drug and pathway association matrix. For these 22 drugs, we obtained their sensitivity spectrum data in 480 cell lines and used the matrix Y3∈R 480×22 For the 58 pathways, we obtained the association data between them and 1802 genes, using the matrix L2∈R22×1802 To record the association data, the number of associations between pathways and genes is 4132. For these 1802 genes, we obtained their expression profile data in 480 cell lines and used the matrix Y4∈R 480×1802 To represent the expression profile of a gene in a cell line.

[0067] 2. Construct a similarity matrix

[0068] 2.1 Construction of similarity matrix based on NCI-60 dataset

[0069] Calculate and construct drug similarity matrix based on NCI-60 dataset Pathway Similarity Matrix and the gene similarity matrix First, we use the Pearson correlation coefficient to calculate and construct the drug similarity matrix based on the drug sensitivity spectrum matrix Y1 in the cell line. The specific calculation method is as follows:

[0070]

[0071] Where nc=57 represents the number of cell lines in the NCI-60 dataset. Since the value range of the Pearson correlation coefficient is [-1,1], we take the absolute value of the calculated result, that is, As the final similarity value between drugs i and j.

[0072] Then, we calculated the pathway similarity matrix based on the pathway and gene association matrix L1 using the Gaussian spectral kernel method as follows:

[0073]

[0074] Among them, np=58 represents the number of pathways in the NCI-60 dataset, Represents the similarity value between paths i and j.

[0075] Then, we calculated the gene similarity matrix based on the gene expression profile matrix Y2 in the cell line using the Pearson correlation coefficient as follows:

[0076]

[0077] Wherein, ng=1863 represents the number of genes in the NCI-60 dataset.

[0078] 2.2 Construction of similarity matrix based on CCLE dataset

[0079] Based on the CCLE dataset, the drug similarity matrix can be constructed using the same method Pathway Similarity Matrix and the gene similarity matrix

[0080] 3. Drug and pathway association prediction method based on cross-layer dependency reasoning in heterogeneous networks

[0081] The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning proposed in the present invention includes four main steps: S1. constructing a three-layer heterogeneous network based on existing data; S2. constructing an optimization model based on the three-layer heterogeneous network; S3. solving the objective function of the constructed model using the block coordinate descent algorithm; S4. predicting the potential association between drugs and pathways based on the result of solving the objective function. The overall process of the method is as follows Figure 1 The following will introduce the content of each step in detail.

[0082] S1. Construct a three-layer heterogeneous network based on existing data. A three-layer heterogeneous network including a drug layer, a pathway layer, and a gene layer will be constructed. First, a drug network layer is constructed. The nodes in the network represent different drugs. If the two drugs corresponding to two nodes are similar, the two nodes are connected by edges, and the weight of the edge is the similarity value of the two drugs. Then, a pathway network layer is constructed. The nodes in the network represent different pathways. If the two pathways corresponding to two nodes are similar, the two nodes are connected by edges, and the weight of the edge is the similarity value of the two pathways. After that, a gene network layer is constructed. The nodes in the network represent different genes. If the two genes corresponding to two nodes are similar, the two nodes are connected by edges, and the weight of the edge is the similarity value of the two genes.

[0083] Then, based on the known drug-pathway association and pathway-gene association data, the nodes in the pathway network layer can be connected to the nodes in the drug network layer and the nodes in the gene network layer, respectively. (It should be noted that based on NCI-60 data and CCLE data, we can construct different three-layer heterogeneous networks. Although the heterogeneous networks are different, the drug-pathway association prediction method proposed by our heterogeneous network cross-layer dependency reasoning is universal.)

[0084] S2. Construct an optimization model based on a three-layer heterogeneous network. Based on the three-layer heterogeneous network of drugs, pathways, and genes constructed in the previous step, we used a regularized optimization model to model the drug and pathway association prediction problem; we constructed the following model to mine the potential feature representation of each layer of the heterogeneous network.

[0085]

[0086] Where i, j represent the i-th and j-th layers in a three-layer heterogeneous network. i, j = 1, 2, and 3 represent the drug layer, pathway layer, and gene layer, respectively. G(i, j) = 1 means that there is an association relationship between the nodes in the i-th and j-th layers. i,j Represents the association matrix between the nodes in the i-th layer and the j-th layer (for example: A 1,2 represents the drug and pathway association matrix). S i is the similarity matrix between the nodes in the i-th layer network (for example, S1 represents the drug similarity matrix). is a weight matrix, n i and n j Represents the number of nodes in layer i and layer j respectively. Matrix T i is the matrix S i The diagonal matrix F i is the low-rank feature matrix of the i-th layer network.

[0087] S3. The objective function of the constructed model is solved by using the block coordinate descent algorithm; by solving the objective function of the above model, the potential feature matrix of each layer of the three-layer heterogeneous network can be obtained. Since the objective function of the model is non-convex, the block coordinate descent algorithm is used to solve the local optimal solution of the objective function. The update rule of the feature matrix of each layer of the heterogeneous network is as follows:

[0088]

[0089] in For nodes u and v, when A i,j (u,v)>0 when I i,j (u,v)=1,A i,j (u,v)=0 when I i,j (u,v)=0. The final F can be obtained by iteratively updating the above formula. i (i=1,2,3). The iterative convergence condition is: reaching the maximum number of iterations t=100 or three characteristic matrices F i The bi-norm of the corresponding difference values ​​of (i=1,2,3) in two iterations is lower than the threshold ξ=10 -8 .

[0090] S4. Predict the potential association between drugs and pathways based on the results of solving the objective function. S4. After solving the potential feature matrix F at the drug layer, pathway layer, and gene layer, i (i=1,2,3), for a given drug i and pathway j, the association score AS(i,j) between them is defined as:

[0091] AS(i,j)=F1(i,:)·F2(j,:) T (11).

[0092] Verification results:

[0093] The prediction performance of the drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning proposed in the present invention was evaluated by performing 100 five-fold cross validations, and the performance was compared with the classic iPaD method in this field based on the AUC index. The results are shown in Table 1:

[0094] Table 1: Results of five-fold cross validation

[0095]

[0096] The heterogeneous network cross-layer dependency inference method of the present invention significantly improves the ability to predict the association between drugs and pathways. By comparing the results, it can be seen that the AUC value obtained by this method is better than the traditional iPaD method on the two sets of data sets. This shows that the method of the present invention can more accurately identify the potential association between drugs and pathways and has stronger classification and distinction capabilities. The present invention utilizes the cross-layer dependency relationship in the heterogeneous network, and this method can better capture the complex correlation between drugs, pathways and other biological entities. Compared with the iPaD method, the heterogeneous network not only considers the single relationship between drugs and pathways, but also comprehensively considers multiple levels of dependency, thereby improving the accuracy and reliability of the prediction results.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the present invention.

Claims

1. A drug and pathway association prediction method based on cross-layer dependency reasoning in heterogeneous networks, characterized in that: The steps include: S1. Build a three-layer heterogeneous network based on existing data; S2. Construct an optimization model based on a three-layer heterogeneous network; S3. Solve the objective function of the constructed model using a block coordinate descent algorithm; S4. Predict the potential association between drugs and pathways based on the results of solving the objective function.

2. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 1 is characterized in that: Step S1 specifically includes: S11. Construct drug layer network, pathway layer network and gene layer network respectively based on the existing data; S12. Construct a three-layer heterogeneous network: Based on the known drug-pathway association and pathway-gene association data, the nodes in the pathway network layer are connected with the nodes in the drug network layer and the nodes in the gene network layer respectively to obtain a three-layer heterogeneous network.

3. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 1 is characterized in that: In step S11, the nodes in the drug network layer network represent different drugs. If the two drugs corresponding to two nodes are similar, the two nodes are connected through an edge, and the weight of the edge is the similarity value of the two drugs; The nodes in the path network layer network represent different paths. If the two paths corresponding to two nodes are similar, the two nodes are connected through an edge, and the weight of the edge is the similarity value of the two paths; The nodes in the gene network layer network represent different genes. If the two genes corresponding to two nodes have similarity, the two nodes are connected through an edge, and the weight of the edge is the similarity value of the two genes.

4. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 3 is characterized in that: In step S11, firstly, a drug similarity matrix S is constructed based on the existing data. d , Pathway Similarity Matrix S p and the gene similarity matrix S g ; Based on the similarity matrix S d , Pathway Similarity Matrix S p and the gene similarity matrix S g Get similarity value; And / or, the existing data include but are not limited to NCI-60 dataset and CCLE dataset.

5. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 4 is characterized in that: The calculation of the drug similarity value comes from the drug similarity matrix S d ; The drug similarity matrix S d The construction is based on the drug sensitivity spectrum matrix Y1 in the cell line, and the Pearson correlation coefficient is used to calculate and construct the drug similarity matrix. The specific calculation method is as follows: Where nc represents the number of cell lines in the data set, and the absolute value of the calculated result is taken, that is, S d (i,j)=|S d (i,j)|, as the final similarity value between drugs i and j.

6. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 4 is characterized in that: The calculation of the similarity value of the path comes from the similarity matrix S of the path p ; The similarity matrix S of the pathway p The construction is based on the pathway and gene association matrix L1, and the Gaussian spectral kernel method is used to calculate the pathway similarity matrix. The specific calculation method is as follows: S p (i,j)=exp(-α p ||L1(i;,)-L1(j;,)|| 2 ) (4) Where np represents the number of pathways in the dataset, S p (i,j) represents the similarity value between paths i and j.

7. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 4 is characterized in that: The calculation of the gene similarity value comes from the gene similarity matrix S g ; The similarity matrix S of the genes g The construction is based on the gene expression spectrum matrix Y2 in the cell line and uses the Pearson correlation coefficient to calculate the gene similarity matrix. The specific calculation method is as follows: Where ng is the number of genes in the dataset; S g (i,j) represents the similarity value between genes i and j.

8. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 1, characterized in that: Step S2. Construct an optimization model based on a three-layer heterogeneous network: Use a regularized optimization model to model the drug and pathway association prediction problem; use the model to mine the potential feature representation of each layer of the heterogeneous network; The objective function of the model is: Among them, i, j represent the i-th and j-th layers in the three-layer heterogeneous network; i, j = 1, 2, and 3 represent the drug layer, pathway layer, and gene layer, respectively; g = 3 represents the number of layers in the network; G(i,j)=1 means that there is an association relationship between the nodes in the i-th layer and the j-th layer; A i,j Represents the association matrix between nodes in the i-th layer and the j-th layer; S i is the similarity matrix between the nodes in the i-th layer network; is a weight matrix, n i and n j Represents the number of nodes in layer i and layer j respectively; Matrix T i is the matrix S i The angle matrix F i is the low-rank feature matrix of the i-th layer network.

9. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 8, characterized in that: In step S3, by solving the objective function, a block coordinate descent algorithm is used to solve the local optimal solution of the objective function; a potential feature matrix of each layer of the three-layer heterogeneous network is obtained; The updating rule of the feature matrix is ​​as follows: in When A i,j (u,v)>0 when I i,j (u,v)=1,A i,j (u,v)=0 when I i,j (u,v)=0; the final F can be obtained by iteratively updating the formula i (i=1,2,3); The iterative convergence condition is: reaching the maximum number of iterations t = 100 or three characteristic matrices F i The bi-norm of the corresponding difference values ​​of (i=1,2,3) in two iterations is lower than the threshold ξ=10 -8 .

10. The drug and pathway association prediction method based on heterogeneous network cross-layer dependency reasoning according to claim 9, characterized in that: In step S4, the potential association between drugs and pathways is predicted based on the result of solving the objective function: after solving the potential feature matrix F at the drug layer, pathway layer, and gene layer, i (i=1,2,3), for a given drug i and pathway j, the association score AS(i,j) between them is defined as: AS(i,j)=F1(i,:)·F2(j,:) T (11)。

Images